Better and Faster Decisions with Recommendation Algorithms

Abstract: While recommendation algorithms have been increasingly used in daily life, little has been done to investigate their effect on decision making in terms of decision quality and preferences. Here we examine this question in an experimental setting whereby subjects from a representative US sample are randomly assigned to five conditions and make sets of binary choices between two lotteries. The two control conditions provide either no recommendation or recommendation based on a randomization device. The three treatment conditions provide recommendations developed by algorithms: one is based on the choice of the majority, and the other two use AI-based recommenders including content-based filtering and user-based collaborative filtering. We find that subjects tend to follow recommended choices and are willing to pay a small fee to receive recommendations for their subsequent decisions. Compared to control conditions, recommendation helps subjects make better and faster decisions and behave in accordance with the independence axiom. These results can be explained by some classes of stochastic choice models. Our work adds to the growing literature on the behavioral underpinnings of algorithms including AI and shed light on the design of choice architecture for decision making under risk.